Comment by dabockster
7 hours ago
You don’t need it if you use llamacpp on Windows, or if you compile it on Linux with CUDA 13 and the correct kernel HMM support, and you’re only using MoE models (which, tbh, you should be doing anyways).
7 hours ago
You don’t need it if you use llamacpp on Windows, or if you compile it on Linux with CUDA 13 and the correct kernel HMM support, and you’re only using MoE models (which, tbh, you should be doing anyways).
What MoE has to do with it? Aside from Flash-MoE that supports exactly one model and only on macOs - you still need to load entire model into memory. You also don't know what experts going to be activated, so it's not like you can predict which needs to be loaded.